NewsNet-SDF: Stochastic Discount Factor Estimation with Pretrained Language Model News Embeddings via Adversarial Networks

TL;DR

This work tackles the challenge of incorporating unstructured textual information into asset pricing by extending the stochastic discount factor (SDF) framework to a multimodal setting. NewsNet-SDF fuses pretrained language model news embeddings, macroeconomic state representations, and firm characteristics through a dedicated fusion module and enforces pricing consistency via an adversarial training mechanism that operationalizes the moment conditions $\mathbb{E}_t[M_{t+1}R^e_{t+1}]=0$. The approach delivers substantial empirical gains, achieving a Sharpe ratio of $2.80$ and reducing pricing errors relative to traditional models by large margins, with ablations showing textual signals contribute more than macro features. The results demonstrate the practical value of integrating semantic news content into asset pricing and risk management applications, including robustness during market disruptions and potential for real-time decision support.

Abstract

Stochastic Discount Factor (SDF) models provide a unified framework for asset pricing and risk assessment, yet traditional formulations struggle to incorporate unstructured textual information. We introduce NewsNet-SDF, a novel deep learning framework that seamlessly integrates pretrained language model embeddings with financial time series through adversarial networks. Our multimodal architecture processes financial news using GTE-multilingual models, extracts temporal patterns from macroeconomic data via LSTM networks, and normalizes firm characteristics, fusing these heterogeneous information sources through an innovative adversarial training mechanism. Our dataset encompasses approximately 2.5 million news articles and 10,000 unique securities, addressing the computational challenges of processing and aligning text data with financial time series. Empirical evaluations on U.S. equity data (1980-2022) demonstrate NewsNet-SDF substantially outperforms alternatives with a Sharpe ratio of 2.80. The model shows a 471% improvement over CAPM, over 200% improvement versus traditional SDF implementations, and a 74% reduction in pricing errors compared to the Fama-French five-factor model. In comprehensive comparisons, our deep learning approach consistently outperforms traditional, modern, and other neural asset pricing models across all key metrics. Ablation studies confirm that text embeddings contribute significantly more to model performance than macroeconomic features, with news-derived principal components ranking among the most influential determinants of SDF dynamics. These results validate the effectiveness of our multimodal deep learning approach in integrating unstructured text with traditional financial data for more accurate asset pricing, providing new insights for digital intelligent decision-making in financial technology.

Figures (6)

• Figure 1: NewsNet-SDF architecture integrating financial news, macroeconomic data, and firm characteristics through specialized pathways into a moment-based SDF framework that enforces the fundamental pricing equation $\mathbb{E}_t[M_{t+1}R_{t+1}^e] = 0$.
• Figure 2: NewsNet-SDF framework architecture: Specialized feature extraction pathways for news, macroeconomic data, and firm characteristics feed into an adversarial network structure that enforces the fundamental asset pricing equation $\mathbb{E}[M_{t+1}R^e_{t+1}]=0$.
• Figure 3: Cumulative Returns of Different Models (1970-2022)
• Figure 4: Top 50 Feature Importance Based on Average Sensitivity
• Figure 5: Cumulative Returns of Beta-Sorted Decile Portfolios (2000-2024)